Get template
From brief to homepage in forty minutes
A founder's first run through Atalier, unedited — every prompt, every revision, and where the agents pushed back.
Shang Chi
Lead Solutions Engineer
We asked a founder who had never used Atalier to build his homepage while we watched, and we agreed in advance to publish whatever happened. No retries, no cherry-picking the good run. He builds inventory software for independent bookshops. He had a name, a rough idea, and no design experience beyond "I know what I don't like." Forty-one minutes later he had a homepage. Here is the transcript, condensed but not cleaned.
Minute 0 to 6 — the brief nobody wants to write
He opened with this: "A homepage for Shelf. It's inventory software for independent bookshops."
That's a description, not a brief, and the Director said so. It asked three questions: who opens this page and in what mood, what happens if they leave, and what would make you personally embarrassed to ship.
The third question is the one that unlocked him. He wrote: "I'd be embarrassed if it looked like every other SaaS page. Bookshop owners are not tech people. They are suspicious of software that looks like it costs a lot."
That sentence did more work than everything before it.
Minute | What happened | Who |
|---|---|---|
0 | Opening brief submitted | Founder |
1 | Three clarifying questions | Director |
6 | Revised brief accepted, spec written | Director |
9 | Three structural directions proposed | Designer |
Minute 9 to 22 — three directions, one argument
The Designer came back with three: a warm editorial layout with a lot of white space, a dense dashboard-forward layout showing the product immediately, and a middle option leading with a single customer quote.
He picked the dense one. The Critic disagreed, and this is where the run got interesting.
Its objection wasn't aesthetic. His own brief had said bookshop owners are suspicious of software that looks expensive. A dense dashboard screenshot above the fold is the single strongest signal of "this is enterprise software" available in web design. He had chosen the layout that contradicted the sentence he'd written eight minutes earlier.
He switched to the editorial direction and kept one dashboard screenshot, moved below the fold, deliberately small.
Minute 22 to 34 — the boring part that matters
Twelve minutes went into things he would never have thought to ask for: a type scale that survives at 320px, a link colour that passes contrast on both the cream and near-black sections, focus states, and a copy pass that cut his hero paragraph from 54 words to 19.
He fought the copy cut. He got 24 words back after arguing, which is roughly the right outcome — the agents don't win every disagreement, and shouldn't.
Where 41 minutes time went |
|---|
Briefing | 6 min |
Direction | 13 min |
Layout build | 12 min |
Copy and polish | 7 min |
Export | 3 min |
Minute 34 to 41 — what shipped
A homepage with a hero, one screenshot, three feature blocks, a pricing table with two tiers, and a footer. No testimonial section, because he doesn't have testimonials. No integrations logos, because he has one integration and a row of one logo looks worse than no row.
Both of those omissions were the agents' suggestion, and both are the kind of restraint that generated pages almost never show. A model optimising for completeness fills every section. A model optimising for the page working leaves gaps where the content isn't there yet.
Where time spent |
|---|
41 min | Brief to finished page |
19 min | Revisions |
2 min | Sections deliberately removed |
1 min | Argument the founder won |
The honest part
It isn't always forty-one minutes. This was a small site with a clear-headed founder who answered questions well. We have runs that take a day because the brief keeps changing, and runs that fail because the founder wants a specific competitor cloned and won't say so directly.
What is consistent is the shape: most of the time goes into deciding, not drawing. The drawing was twelve minutes. The deciding was twenty-nine. That ratio holds almost everywhere, which suggests the useful thing agents automate isn't the design work — it's the waiting between design work.